Re: [RFC PATCH 00/52] Make GLIBC Y2038-proof

[Joseph Myers <joseph at codesourcery dot com>]

On Fri, 8 Sep 2017, Zack Weinberg wrote:

> > And as discussed, I disagree and don't think we should invent any such
> > typedef or deviate from this requirement, given that long is fully able to
> > represent all valid nanoseconds values
> 
> Neither you, nor Rich, nor anyone else on the cited Austin Group
> discussion has addressed the actual issue, which is that this is a
> datum embedded in structures passed across the kernel/user boundary by
> reference, it is impossible to enumerate all such structures,
> therefore its type _must_ be a typedef so that it can be made to match
> the kernel's expectations, whatever those might be.

I don't see this as in any way an issue for having a definition of struct 
timespec that conforms to ISO C and POSIX requirements.  Code passing data 
to the kernel knows the structures being used locally and needs to convert 
it to the kernel's layout, which might be e.g. struct __kernel_timespec64.  
I don't see that as different in essence from e.g. converting a local-time 
broken down date/hours/minutes/seconds/nanoseconds value before passing it 
to the kernel, if that's the form the particular application stores time 
in.  It just so happens that, as an optimization, we can arrange the 
layout so that we both are POSIX-compatible and do not require any layout 
conversions between timestamps coming *from* the kernel and struct 
timespec.

> (And I don't particularly see why people seem to think this is an
> x32-specific issue.  It potentially affects *any* 32-bit ABI on a
> 64-bit kernel.)

The kernel does the conversion just fine for a pure 32-bit struct timespec 
- as it does for the entire compat interface - and could do them perfectly 
well for a structure with 64-bit seconds and 32-bit (+ padding) 
nanoseconds.  It just so happens not to for the very specific case of x32.

> >> <sys/types.h>, matching the kernel's choice of 32 or 64 bits for this
> >> field (all _t names are reserved for future extension, so the typedef
> >> doesn't need to be _GNU_SOURCE-only). This "willful violation" (as the
> >
> > They are not reserved in ISO C, only in POSIX.
> 
> <sys/types.h> doesn't exist in ISO C as far as I know, so how could it
> be?  But any program that uses it opts into POSIX's specification for

My point is that you can't expose nsec_t, even as a typedef for long, in 
<time.h> for C11.

-- 
Joseph S. Myers
joseph@codesourcery.com